P R ! F Y S G O L
CArRDv^
BINDING SERVICES T e l+44 (0)29 2087 4949 Fax +44 (0)29 20371921 e-mail [email protected]Computerised Electrocardiogram
Classification
A thesis submitted to the University of Wales, Cardiff in candidature for the degree of
Doctor of Philosophy
by
Zakria Zaki Mahrousa, B.Eng.
Manufacturing Engineering Centre School of Engineering
Cardiff University United Kingdom
All rights reserved
INFORMATION TO ALL USERS
The quality of this reproduction is dependent upon the quality of the copy submitted.
In the unlikely event that the author did not send a complete manuscript and there are missing pages, these will be noted. Also, if material had to be removed,
a note will indicate the deletion.
Dissertation Publishing
UMI U584661
Published by ProQuest LLC 2013. Copyright in the Dissertation held by the Author. Microform Edition © ProQuest LLC.
All rights reserved. This work is protected against unauthorized copying under Title 17, United States Code.
ProQuest LLC
789 East Eisenhower Parkway P.O. Box 1346
SYNOPSIS
Advances in computing have resulted in many engineering processes being automated. Electrocardiogram (ECG) classification is one such process. The analysis of ECGs can benefit from the wide availability and power of modem computers.
This study presents the usage of computer technology in the field of computerised ECG classification. Computerised electrocardiogram classification can help to reduce healthcare costs by enabling suitably equipped general practitioners to refer to hospital only those people with serious heart problems. Computerised ECG classification can also be very useful in shortening hospital waiting lists and saving life by discovering heart diseases early.
The thesis investigates the automatic classification of ECGs into different disease categories using Artificial Intelligence (AI) techniques. A comparison of the use of different feature sets and AI classifiers is presented. The feature sets include conventional cardiological features, as well as features taken directly from time domain samples of an ECG. The benchmark AI classifiers tested include those based on neural network, k-Nearest Neighbour and inductive learning techniques.
The research proposes two modifications to the learning vector quantisation (LVQ) neural network, namely the All Weights Updating-LVQ (AWU-LVQ) algorithm and
the Neighbouring Weights Updating-LVQ (NWU-LVQ) algorithm, yielding an “intelligent” diagnostic heart system with higher accuracy and reduced training time compared to existing AI techniques.
To my parents,
my wife %aw fa La6a6idi
and my daughters Sara and Sidra,
without their (oving support
this would not have happened
Acknowledgements
First of all I thank Allah (my Lord) the all high, the all great who made it possible for me to complete this work.
I would like to thank my supervisor Professor D. T. Pham for his excellent supervision, continuous encouragement and support. He was a brilliant supervisor. He really deserves more thanks than I can properly express for the great deal o f time he gave me, and for making that time the most enjoyable and rewarding hours of my academic life.
I would like to thank Aleppo University for sponsoring me during my Ph.D. studies.
I reserve my deepest gratitude for my parents, Mohammed Zaki Mahrousa and
Fawzia Rehawai who gave and have given me continuous support and encouragement throughout my studies, and my parents-in-law M. Amine Lababidi
and Naila Kouka for their encouragement. Many thanks also to my brothers, sisters, brothers-in-law, sisters-in-law and a special thanks goes to brother M. Yehaya Mahrousa for his brotherhood advice.
My thanks are also due to all those members of the Intelligent System Laboratory who contributed to my work through their valuable help, advice and technical support. In particular, I would like to mention Dr.E. E. Eldukhri.
more than any technical help. Special thanks also go to Dr. Y. Dadam, Dr. Z. Salem, Dr. B. AI Mourad, Mr. S. Otri, Mr. M. R. Assi and all those who made my stay in the laboratory particularly pleasant.
Declaration
This work has not previously been accepted in substance for any degree and is not being concurrently submitted in candidature for any degree.
Date...
Statement 1
This thesis is the result of my own investigations, except where otherwise stated. Other sources are acknowledged by footnotes giving explicit references. A bibliography is appended.
Date...
Statement 2
I hereby give consent for my thesis, if accepted, to be available for photocopying and for inter-library loan, and for the title and summary to be made available to outside organisations.
Signed (Zakria Mahrousa - Candidate)
Signed (Zakria Mahrousa - Candidate)
Signed Date...
(Zakria Mahrousa - Candidate)
Contents
Synopsis iii Dedication v Acknowledgem ents vi D eclaration viii Contents ixList o f Figures xiii
List o f Tables xvii
Abbreviations xix
Nomenclature xxi
Chapter 1 - Introduction 1
1 . 1 Computerised Electrocardiogram Classification 1
1.2 Research Topic 3
1.3 Objectives 4
1.4 Outline of the Thesis 5
Chapter 2 - Review of Automated ECG Classification 7
2.1 Preliminaries 7
2.2 The Heart 8
2.2.1 Electricity in the Heart 11
2.2.2 Electrocardiogram 13
2.2.3 Recording ECGs 13
2.2.4 ECG Wave-Form 14
2.3 Cardiac Arrhythmia 19
2.3.1 MIT-BIH Arrhythmia Database 19
2.4 Data Preparation 35
2.4.1 Feature Extraction 35
2.4.2 Normalisation 41
2.5 Previous W ork on ECG/Arrhythmia Classification 51
2.6 Summary 58
Chapter 3 - Comparison of Different ECG Classification Techniques 59
3.1 Pattern Classification 59
3.2 Instance-based Learning Classifiers 64
3.3 Artificial Neural Networks 67
3.3.1 Neural Network Structure 6 8
3.3.2 Learning in Neural Networks 69
3.4 Multilayer Perceptron 71
3.6 Inductive Learning 8 8
3.7 Simulation Results and Comparisons 96
3.8 Summary 104
Chapter 4 - Enhanced Learning Vector Quantisation Network using All Weights
Updating 105
4.1 Preliminaries 105
4.2 Learning Vector Quantisation (LVQ) 106
4.3 All Weights Updating LVQ (AWU-LVQ) 110
4.4 Experimental Results 115
4.4.1 Network Configuration 115
4.4.2 Training and Test Sets 117
4.4.3 AWU-LVQ and LVQ Results 118
4.5 Summary 125
Chapter 5 - Enhanced Learning Vector Quantisation Network using
Neighbouring Weights Updating 126
5.1 Preliminaries 126
5.2 Self Organising Map (SOM) 127
5.3 Fuzzy Soft Learning Vector Quantisation (FS-LVQ) 131
5.5 Experimental Results 134
5.5.1. Network Configuration 135
5. 5.2 Training and Test Sets 135
5.5.3 NWU-LVQ and FS-LVQ Results 137
5.4 Summary 143
Chapter 6 - Conclusions and Future Work
144
6.1 Preliminaries 144
6.2 Conclusions 144
6.3 Contributions 146
6.4 Future Work 146
Appendix A: Training and Test Data Extracted from the MIT-BIH Database
148
Appendix B: Statistical Formulae used in Feature Selection 155
List of Figures
Chapter 1
Figure 1.1: Three processing modules of the computerised ECG system
Chapter 2
Figure 2.1: Cutaway drawing of the human heart showing the chambers, valves and connecting blood vessels (adapted from [ Texas Heart Institute,
2002])
Figure 2.2: Electrical conduction pathways within the heart (adapted from [Wahl, 1999])
Figure 2.3: The twelve standard lead positions for recording the ECG (adapted from [www.biopac.com])
Figure 2.4: Idealised representation of an ECG trace (lead II) for one normal heartbeat
Figure 2.5: Normal sinus rhythm (N) type (MIT-BIH databases, record 100)
Figure 2.6: Left bundle branch block (L) type (MIT-BIH database, record 109)
Figure 2.7: Right bundle branch block (R) type (MIT-BIH database, record 118)
Figure 2.8: Beat stimulated by an artificial pacemaker ( ‘Paced’) (MIT-BIH database, Record 104)
2 9 12 15 17 23 24 25 27 xiii
Figure 2.9: Premature ventricular contraction (V) type (MIT-BIH database, record 105)
Figure 2.10: Atrial premature beat (A) type (MIT-BIH database, record
100)
Figure 2.11: Aberrated atrial premature beat (a) type (MIT-BIH database, record 105)
Figure 2.12: Nodal (junctional) escape beat (j): type (MIT-BIH database, record 2 0 1)
Figure 2.13: Ventricular escape beat (E) type (MIT-BIH database, record 207)
Figure 2.14: Fusion of paced and normal beats beat (f) type (MIT-BIH database, record 113)
Figure 2.15: Some features extracted from an ECG
Figure 2.16 (a): Variations in the R-wave part of QRS com plex showing inverted R-wave
Figure 2.16 (b): Variations in the R-wave part o f QRS com plex showing S-wave has greater magnitude than R-Wave
Figure 2.16 (c): Variations in the R-wave part of QRS com plex showing apparent double R peak
Figure 2.17 (a): Variations in the Q-wave part of QRS complex showing depressed PR segment
Figure 2.17 (b): Variations in the Q-wave part of QRS complex showing no Q-wave present
indeterminate onset to Q-wave 47 Figure 2.18 (a): Variations in the S-wave part of QRS complex showing
indeterminate end to S-wave 48
Figure 2.18 (b): Variations in the S-wave part of QRS complex showing
depressed ST segment 49
Figure 2.18 (b): Variations in the S-wave part of QRS complex showing
apparent separation of S-wave from QR Part of complex 50
Figure 2.19: RR intervals from an ECG 55
Chapter 3
Figure 3.1: Components of an ECG pattern classification system 62
Figure 3.2: Machine learning techniques classification 63
Figure 3.3: An example of k-NN classification 65
Figure 3.4: Simplified representation of supervised learning for NNs 70 Figure 3.5: Simplified representation of unsupervised learning for NNs 72
Figure 3.6: A neuron in a multilayer perceptron 74
Figure 3.7: A fully connected multilayer perceptron (MLP) 75
Figure 3.8: Back propagation multilayer perceptron 78
Figure 3.9: Radial basis function network 84
Figure 3.10: A simple decision tree example 90
Figure 3.11: Example of a decision tree extracted by C5.0 94
Figure 3.12: Graphical representation of the data in Table 3.3 100 Figure 3.13: Effect of removing different features of the C5.0 on the classification
accuracy 103
Chapter 4
Figure 4.1: Learning vector quantisation network 107
Figure 4.2: Gaussian weight updating function 112
Figure 4.3 (a): Exponential decay function 114
Figure 4.3 (b): Linear decay function 114
Chapter 5
Figure 5.1: A typical two-dimensional Kohonen self organising map 128
List of Tables
Chapter 2
Table 2.1: The twelve standard leads positions used in ECG 16
Table 2.2: Summarised values of different intervals and segments of a normal ECG
signal with different heart rates 2 0
Table 2.3: 18 Features of the ECG signal selected for classification 39 Table 2.4: 11 Features of the ECG signal selected for classification 40
Chapter 3
Table 3.1: Activation functions [Pham and Liu, 1995] 76
Table 3.2: Number of training and test examples taken from the MIT database 98 Table 3.3: Summary of the classification accuracy using different classification
techniques 99
Table 3.4: Effect of removing features on the classification results using C5.0 102
Chapter 4
Table 4.1: Features of the ECG signal selected for classification 117 Table 4.2: Effect of kernel width on the classification accuracy
(learning rate a = 0 . 1 and epoch number = 200) 119
Table 4.3: Effect of epoch numbers on the classification accuracy
(learning rate a = 0.1 and kernel width a = 2) 120
Table 4.4: Effect of epochs numbers on AWU-LVQ classification accuracy
(learning rate a = 0 . 1 and kernel width a = 2) 1 2 2
Table 4.5: LVQ Classification results (learning rate a = 0.1) 123
Table 4.6: Comparison between AWU-LVQ and LVQ 124
Chapter 5
Table 5.1: NWU-LVQ classification accuracy with different numbers of training epochs (learning rate a = 0.1 and kernel width a = 3) 136 Table 5.2: NWA-LVQ classification accuracy with different kernel values with
(learning rate a = 0.1 and epoch number = 1 0 0) 138
Table 5.3: Classification accuracy of NWU-LVQ for different epoch number
(learning rate a = 0.1 and kernel width a = 3) 140 Table 5.4: Classification accuracy of FS-LVQ results for different epoch numbers
(learning rate a = 0 . 1 and kernel width a = 3) 141
Table 5.5: Summary of classification results 142
Abbreviations
a Aberrated Atrial Premature Beat
A Atrial Premature Beat
AI Artificial Intelligence
ANN Artifical Neural Network
ART Adaptive Resonance Theory
A V Atr io V entricular
AWU-LVQ All Weights Updating - LVQ
BIH Beth Israel Hospital
BP Back Propagation
E Ventricular Escape Beat
ECG Electrocardiogram
EEG Electro Encephalogram
F Fusion of paced and normal beat
FS-LVQ Fuzzy Soft-Learning Vector Quantisation
j Nodal (junctional) Escape Beat
k-NN k-Nearest Neighbour
L Left Bundle Branch Block Beat
LA Left Arm wrist
LL Left Leg ankle
LVQ Learning Vector Quantisation
, MIT-BIH Massachusetts Institute of Technology- Beth Israel Hospital
MLP Multi-Layer Perceptron
N Normal Beat
NN Neural network
NWU-LVQ Neighbouring Weights Updating - LVQ
P Paced Beat
R Right Bundle Branch Block Beat
RA Right Arm wrist
RBF Radial Basis Function
RL Right Leg ankle
SA Sino Atrial
SLPs Single-layered Perceptrons
SOFM Self Organising Feature Map
SOM Self Organising Map
SVMs Support Vector Machines
V Premature Ventricular Contraction
ZISC Zero Instruction Set Computer
NOMENCLATURE
Chapter 2
F s F 1 mm f 1 max Fu X d>Chapter 3
xu
x(() ni K l (k,n) Class (x) D{x,x[i)) d M ( x , x (i)) yjJO)
Aw..Scaled value after normalisation Minimum value in the data set Maximum value in the data set Unsealed value before normalisation Mean of the data set
Standard deviation of the data set
Unlabeled sample Training pattern vector
Number of training pattern from class I
Number of patterns from class among the k-NN of pattern x k-NN decision rule
Euclidean distance between two pattern vectors x and x(l)
Number of features
Matching score for the test pattern x and the training pattern xl
Output of the neuron j
Activation function
Weights adjustment between neurons i and j in the forward direction
rj Learning rate
5 j Error factor of neuron j
netj Total of all the inputs to neuron j dj Desired output of neuron j
ojp Gaussian function
a] Variance of the Gaussian function
x ^ Input vector for index pattern p
r Momentom convergence
k Output unit index
K Total number of output units
y kp Output of the RBF network
P Total number of patterns
Pkp Degree to which pattern p belongs to class k
t Time step
tp Target for pattern p
w ki Weight from the i th hidden unit to the k th output Awki Weight updates for the winning output unit ykp
Discrete attribute A with n possible values
Q The ith class of a labelled data set S with k classes P(Ci,S) proportion of instances in S which are in class
C,-S labelled data set
Ent(S) Entropy of a labelled data set S
Chapter 4
di Euclidean distance between the input vector x and the weight vector w,
dk Euclidean distance between the next closest neuron of the winning neuron and the input vector
Wy 7th components of the weight
Xj 7th components of the input vector x
x Input vector
w Weight vector
a Learning rate
h Neighbourhood function
g Width of the kernel of the Gaussian function
P A parameter used to eliminate negative corrections in supervised learning
Chapter 5
x Input vector
w Weight vector
a Learning rate
h Neighbourhood function of the winner j r j, rk Vectorial locations for the display grid a Width of the kernel of the Gaussian function
a (0) Starting value of the learning rate
a 0 Rate of decrease
// Fuzzy C Mean (FCM) membership functions
fit) Function of t that controls the approximated excitation of the FS-LVQ neurons
h Neighbourhood function
p A parameter used to eliminate negative corrections in supervised learning
Chapter 1
Introduction
1.1 Computerised Electrocardiogram Classification
Since the advent of computerised ECG classification systems in 1957 [Titomir, 2000], research in the field has proliferated. Each year, increasing numbers of ECGs are recorded using ECG recording systems of on-board ECG diagnosis facilities, for example to detect significant changes in the patients’ physiological state. The aim of ECG classification is to determine if the patient is ill and requires treatment or if the patient has no cardiac abnormalities and requires no treatment. In general, computerised ECG systems can assist the non-specialist with patient diagnosis, offering greater consistency through avoidance of observer variability. A computerised system normally consists of three processing modules: beat detection, feature extraction and classification, as shown in Figure 1.1. Beat detection concerns the identification and location of each cardiac cycle. The reference markers generated by the beat detection module are processed by the feature extraction module to produce a feature vector describing the morphology of the recorded ECG. The feature vector is then examined by the classification module, giving a hypothesis based on the available evidence.
ECG
Feature Extraction Beat
Detection Classification
Figure 1.1: Three processing modules of the computerised ECG system
ECG
* Class
some extent reached a plateau with regard to achieved performance levels. In order to enhance the overall performance of the system so that it could be used reliably, the classification stage is therefore now being targeted
1.2 Research Topic
This research concerns the automatic classification of ECGs for arrhythmia analysis. A comparison was carried out of different classification techniques drawn mainly from the field of Artificial Intelligence (AI). AI classification techniques were considered as they can learn the classification task automatically. They do not require a priori modelling and are simpler to apply, yet promise to yield good performance.
The work focused on developing improved AI classification tools for distinguishing between abnormal ECGs (indicating the presence of arrhythmias) and normal ECGs (showing no signs of arrhythmias).
The AI techniques investigated in this research were neural networks and inductive learning. Neural networks were focused upon due to their superior abilities in handling continuous inputs such as those derived from an ECG. Different neural networks were studied and improvements were made to the Learning Vector Quantisation (LVQ) network, a well-known classification tool.
1.3 Objectives
The main objectives of this research were:
• To compare the performance of AI techniques such as neural networks and inductive learning for computerised ECG classification. The k-Nearest Neighbour method was also considered in the comparison.
• To compare the classification performance using different features extracted from ECG signals and different signal sampling rates. The aim was to determine the features and sampling rates that characterise ECG signals well and can be used successfully for classification purposes.
• To develop improved Learning Vector Quantisation (LVQ) algorithms in order to achieve better classification accuracy with a short training time. Two modifications were made to the standard LVQ algorithm to yield new LVQ algorithms. The first modification to the LVQ algorithm is called the All Weights Updating LVQ (AWU-LVQ). This gave better classification accuracy with a short training time compared to the standard LVQ algorithm. The second modification is called the Neighbouring Weights Updating LVQ (NWU-LVQ). This has a higher classification accuracy and required a shorter training time compared to other LVQ Algorithms.
1.4 Outline of the Thesis
This thesis comprises six chapters and two appendix.
Chapter 2
reviews the background literature relevant to the work presented in the thesis.Chapter 3
presents a comparison of different classification techniques including decision tree inductive learner, k-Nearest Neighbour classifier, multilayer perceptrons and radial basis function neural networks.Chapter 4
describes an enhancement of the Learning Vector Quantisation technique called All Weights Updating.Chapter 5
describes another enhancement of the Learning Vector Quantisation technique called Neighbouring Weights Updating.Chapter 6
presents the conclusions of the research and recommendations for further study.Appendix A details the examples data extracted from the Massachusetts Institute of Technology-Beth Israel Hospital (MIT-BIH) database as used in this study.
Appendix B Statistical Formulae used in Feature Selection.
Chapter 2
Review of Automated ECG Classification
2.1 Preliminaries
An electrocardiogram (ECG) is an important tool in clinical diagnosis relating to disorders of the human heart. It records the electrical activity of the heart in terms of voltage changes transmitted to the body surface by electrical events in the heart.
Automatic ECG beat detection and classification is an essential tool for clinical settings, such as an intensive care unit. Such systems, however, must be able accurately to detect and classify problems on a real time basis. It is known that several arrhythmias are potentially dangerous and even life threatening if not detected within a few seconds of their onset.
One major problem facing today’s automatic ECG analysis equipment is the wide variation in the morphologies of ECG waveforms of different patients and patient groups.
This chapter gives a review of the physiology and functionality of the human heart and how ECG signals are measured including the placement of the leads, and the nature and characteristics of ECG signals. The Chapter also provides a description of
the database used in this work and a literature review on ECG arrhythmia classification.
2.2 The Heart
The heart is a muscle weighing approximately 300 grams. It consists mainly of two types of cells, those which form the working muscle, and those which control electrical signals. Its main function is to maintain the supply of blood in the body and to deliver it to all parts of the body via the arteries. In order to achieve this, the heart muscle must pump approximately 7200 litres of blood per day. At rest, the normal heart beats around 72 times and pumps around 5 litres of blood per minute. According to the demands of the body, it is capable of rapidly increasing this to 15 litres per minute with a heart rate of up to 2 0 0 beats per minute or more.
Figure 2.1 shows the physical structure of the heart. The heart consists of two chambers [McCulloch and Bastion, 1999], which act as two synchronised pump stages. The first chamber comprises the atria, which consist of the left and right atria. The second chamber is made up of the ventricles, which consist of the left and right ventricles. The right atrium and right ventricle supply blood to the lungs for oxygenation (pulmonary circulation), whereas the left atrium and left ventricle supply blood to the rest of the body (systemic circulation).
Excitation of the heart does not come directly from the central nervous system but is initiated in the SinoAtrial (SA) node (see Figure 2.2), or pacemaker, which are a special group of excitable cells.
V e n a C ava Aorta Pulmonary Artery Left 0 Atrium Right Atrium V felve Tricuspid
V&lve I ! v&jve Left \ yentricle,-' Right
Ventricle
Figure 2.1: Cutaway drawing of the human heart showing the chambers, valves and connecting blood vessels (adapted from [Texas Heart Institute, 2001])
Spread o f "depolarisation B u n d le o f H is Left atrium Right Atrium Right \ ventricle v. ventricle A V N o d e R ight B u n d le Branch Purkinje Fibres SA N o d e
Figure 2.2: Electrical conduction pathways within the heart (adapted from [Wahl, 1999])
the rate is influenced by nerves that accelerate or inhibit its value.
To initiate the heartbeat, the action potential generated by the pacemaker propagates (from the SA node) in all directions along the surface of both atria towards the junction of the atria and ventricles (the AV node - see Figure 2.2). At this point, special nerve fibres act to delay the propagation to provide proper timing between the pumping action of the atria and the ventricles. During this delay time, the atria complete their contraction, forcing blood into the ventricles in order to complete their filling. Afterwards the AV node initiates an impulse into the ventricles and then into the bundle branches that connect to special cardiac muscle fibres (the Purkinje fibres) in the myocardium The latter is the muscular middle layer of the wall of the heart. It is composed of spontaneously contracting cardiac muscle fibres which allow the heart to contract.
The front wave in the ventricles, however, does not follow along the surface but is perpendicular to it and moves from the inside to the outside of the ventricular wall, until the entire ventricle becomes depolarised. Then the ventricles contract, forcing blood into the pulmonary and systemic circulatory systems. A wave of depolarisation follows the repolarisation wave by about 0.2 to 0.4 second. This depolarisation, however, is not initiated from neighbouring muscle cells but occurs as each cell returns to its resting potential independently.
2.2.1 Electricity in the Heart
The cardiac muscle contracts when it is stimulated by tiny electrical impulses carried in conductive fibres within the muscle. Unlike other muscles, the heart is unique. It generates its own electricity rather than receiving signals initiating in the brain and central nervous system This allows the most important muscle in the body to function independently. In a normal healthy heart, these electrical impulses, of the order of 1 mV, originate in the SA node and follow a pathway around the heart in the conductive fibres. Conduction occurs due to potassium and sodium ions passing through cell walls [McCulloch and Bastion, 1999]. Figure 2.2 shows the electrical pathways within the heart.
To measure the heart’s electrical activity, small metal leads are used. When a lead makes contact with the surface of the skin, which acts as an electrolyte, a difference of potential between the lead and the skin is produced.
The top layer of skin consists largely of dead cells along with a certain amount of oil and grime. Therefore, the skin’s natural electrical resistance is high compared to the resistance of the body’s fluids. An electrolytic jelly or paste is usually applied between the lead and the skin to ensure a low value of lead-skin interface resistance and minimise the drop in potential across the interface. Larger surface leads tend to have lower resistance compared to small needle leads.
2.2.2 Electrocardiogram
An ECG records the electrical activity of the heart, in particular the propagation of the electric potential through the cardiac muscles, as monitored by sensors at the limb extremities. It is considered a representative signal of the cardiac physiology, useful in diagnosing cardiac disorder. The state of cardiac health is generally reflected in the shape of the ECG waveform and heart rate. It may contain important pointers to the nature of diseases afflicting the heart [Acharya et al., 2003]. The biopotentials generated by the muscles of the heart result in the ECG.
2.2.3 Recording ECGs
The first accurate recording of an ECG was made by Einthoven in 1895 using a string galvanometer [Jenkins, 2001]. In order to record the ECG from electrodes placed vertically as well as horizontally on the human body, he had the electrodes placed not only on the chest but also in both arms and one leg. The leg selected was the left one, probably because it terminates vertically below the heart. Nowadays, an ECG is recorded by affixing up to five electrodes to the body [DeMarre and Michaels, 1983]. Usually two electrodes, or one electrode plus a group of up to three others, are connected to an amplifier. The electronic amplifier requires an additional connection to the body as a ground reference. The free right leg is used for this purpose. The lead placements are the right arm wrist (RA), left arm wrist (LA), left leg ankle (LL), right leg ankle (RL) and chest. As for the chest leads, different positions have been used.
Figure 2.3 shows the twelve standard lead positions over the body, and Table 2.1 lists the connections between the leads and their types (bipolar, and unipolar limb leads, unipolar chest leads).
2.2.4 ECG Wave-Form
The ECG wave-form for one cardiac cycle is shown in Figure 2.4. The letter designations given to each of the prominent features are those conventionally adapted.
The waveform is interpreted as follows [Hampton, 1998]:
The P-wave indicates the SA node function, which is produced by atrial depolarisation. During atrial depolarisation, the potential’s action travels from the SA node towards the atrioventricular (AV) node.
The R-wave represents the depolarisation of most (but not quite all) of the remaining ventricular musculature. Because the ventricular muscle is massive compared to the atrial muscle, the R-wave amplitude is much higher than the P-wave amplitude. The R-wave, like the P-wave, appears above the baseline and is usually the most prominent feature in the ECG. The normal peak value of the R-wave is approximately 1 mV when measured at the surface of the body and about 40 mV when measured inside the heart.
II
Figure 2.3: The twelve standard lead positions for recording the ECG (adapted from [www.biopac.com])
Bipolar Limb Leads
Lead I LA (+input) & RA (- input) Lead II LL (+input) & RA (- input) Lead III LL (+input) & LA (- input) Unipolar Limb Leads
Lead aVR RA (+input) & LA+LL (-input) Lead aVL LA (+input) & RA+LL (-input) Lead aVF LL (+input) & RA+LA (-input) Unipolar Chest Leads
VI Chest (+input) & RA+LA+LL (-input) V2 Chest (+input) & RA+LA+LL (-input) V3 Chest (+input) & RA+LA+LL (-input) V4 Chest (+input) & RA+LA+LL (-input) V5 Chest (+input) & RA+LA+LL (-input) V6 Chest (+input) & RA+LA+LL (-input)
RA = right arm wrist, LA = left arm wrist, LL = left leg ankle + input refers to the positive terminal of the ECG recorder - input refers to the negative terminal of the ECG recorder
Table 2.1: The twelve standard lead positions used in ECG
potential /M illi- Volt (mV) 1.0 0 .5 0.0 -0 .5 0.0 0.2 0 .4 0.6 0.8.8 0 Time/Second
Figure 2.4: Idealised representation of an ECG trace (lead II) for one normal heartbeat
The PR interval indicates AV conduction time. The interval is measured from the onset of the P-wave to the beginning of the QRS complex. A normal PR interval is 0.18 to 0.20 seconds. A short PR interval indicates that the impulse originates in an area other than the SA node, and a long PR interval indicates that the impulse is delayed as it passes through the AV node. Whereas atrial depolarisation occurs in one direction, the ventricles depolarise in three directions. Immediately after the impulse delay period, initial depolarisation of the ventricles begins in the septal area just below the AV node. Because the walls of the left ventricle are thicker than the walls of the right ventricle, the depolarising wave travels from left to right, causing the left side of the body to become negative while the right side becomes positive. This is recorded on the Q-wave, which appears below the baseline. Normally the amplitude of the Q-wave is less than the amplitude of the P-wave and in some tracings the Q- wave is not seen.
The S-wave represents the depolarisation of the remaining portion of the ventricles. Since for this wave, the apex (see Figure 2.1) becomes negative while the AV node area becomes positive, the recorded S-wave appears below the baseline. In general, the amplitude of the S-wave is greater than that of the Q-wave.
However, for some patients, the S-wave amplitude is so small that it is not observable. The QRS complex is the combined result of the repolarisation of the ventricles. Prior to this time interval, the atria are depolarising. However, because of its small amplitude, the depolarising wave pattern of the atria is not measurable at the body’s surface. Thus the baseline normally remains flat between the end of the P-wave and the start of the QRS complex.
depolarisation is represented by the T-wave.
Occasionally, another wave will appear after the T-wave. This is called the U-wave,
and is generally believed to be the result of after potentials in the ventricular muscle.
The U-wave is more frequently seen in tracings from infants and patients who have low serum potassium levels or an enlarged heart. Following depolarisation, the ventricles relax.
Table 2.2 summarises the values of different intervals and segments of a normal ECG signal with different heart rates. This table shows that the heart rate affects the ECG signals even if the person is healthy.
2.3 Cardiac Arrhythmia
Arrhythmias, also known as dysrhythmias, refer to any disorder of heart rate or rhythm In a normal heart, the atria and ventricles contract in a co-ordinated manner. The depolarisation wave spreads from the SA node, through the atria, to the AV node and through the bundle of His (see Figure 2.2) and bundle branches into the ventricles. A conduction abnormality can occur at any of these points.
2.3.1 MIT-BIH Arrhythmia Database
The Massachusetts Institute of Technology Beth Israel Hospital (MIT-BIH) arrhythmia database [Moody and Darker, 1989] is a well-established source of
P-R Interval QRS interval Rate Q-T Interval S-T Segment 0.18 to 0.20 Second 0.07 to 0.01 Second 60 70 80 90 1 0 0 1 2 0 0.33 to 0.43 Second 0.31 to 0.41 Second 0.29 to 0.38 Second 0.28 to 0.36 Second 0.27 to 0.35 Second 0.25 to 0.32 Second 0.14 to 0.16 Second 0.13 to 0.15 Second 0.12 to 0.14 Second 0.11 to 0.13 Second 0.10 to 0.11 Second 0.06 to 0.07 Second
Table 2.2: Summarised values of different intervals and segments of the normal ECG signal with different heart rates
ECG signals, which have been transferred from Holter tape recordings taken from various in-patients at the Arrhythmia hospital laboratory at the Beth Israel Hospital between 1975 and 1979. From 4000 Holter tape records, 48 annotated records divided into two groups were kept. Group one consists of 23 records (the lxx series) and contains examples that an arrhythmia detector might encounter in routine clinical use. The second group consists of 25 records (the 2xx series) and contains examples of complex arrhythmias that could pose difficulties to arrhythmia detectors or of rare clinical cases. The subjects were 25 men aged 32 to 89 years, and 22 women aged 22 to 89 years. About 60% of the records were obtained from in-patients. Each record is slightly over 30 minutes in length. The signals were sampled at the same frequency, 360 Hertz, but not necessarily at the same gain because during collection different equipment was used with different electrical gains for digitisation of the various records. Moreover, the digital amplitude values range between [0, 2047], where 1024 represents 0 volts. Therefore, the signals require normalisation before use.
The variety of the patients and variation in their ages and physical conditions makes the MIT-BIH database suitable for investigations into ECG classification techniques.
Lead II was the lead type used to record most of the ECG signals in the MIT-BIH database.
The MIT-BIH Arrhythmia database contains software to enable extraction of the digitised records. For the purpose of this study the following ECG types were selected from the MIT-BIH database:
1- Normal Sinus Rhythm (N): this is the term for the normal condition (see Figure 2.5).
2- Left Bundle Branch Block Beat (L): this arrhythmia is caused by a problem in conduction in the His bundle in the left side ventricle. This is seen as a widening of the QRS complex. This ECG type is invariably an indication of heart disease [Hampton, 1998]. Figure 2.6 indicates that the QRS complex is notably wider than that shown in Figure 2.5. This is due to the extra time taken for depolarisation caused by poor electrical conduction (block).
3- Right Bundle Branch Block Beat (R): the cause of this arrhythmia is similar to
(L). However, the conduction problem now occurs on the right side of the His bundle branch and the ECG indicates a problem in the heart but also can be seen in a healthy heart. This type of arrhythmia is identified by a wide bimodal QRS complex (see Figure 2.7).
4- Paced Beat (P): this problem arises in patients that have been fitted with an artificial pacemaker. Pacemakers are used when a person has bradycardia (a very slow heart rhythm), which causes poor circulation and cannot be corrected by treatment with drugs. Pacemakers stimulate the heart muscle. This type of arrhythmia is indicated by the occasional missing of the P-wave and the presence of a spike
potential (mV) 0 0 .5 0 -0 .5 0 0.2 0 .4 0.6 0.8 Time (seconds)
Figure 2.5: Normal sinus rhythm (N) Type (MIT-BIH database, record 100)
Electrical potential (mV) 0 .7 5 0 .5 0 0 .2 5 -0 .2 5 -0 .5 0 -0 .7 5 0.6 0 .4 0.2 0 Time (seconds)
Figure 2.6: Left bundle branch block (L) type
(MIT-BIH database, record 109)
potential (mV) 0 .7 5 0 .5 0 0 .2 5 -0 .2 5 -0 .5 0 -0 .7 5 0 0.2 0 .4 0.6 Time (seconds)
Figure 2.7: Right bundle branch block (R) type
representing the stimulus from the pacemaker, followed by a wide QRS complex (see Figure 2.8).
5- Premature Ventricular Contraction (V): this arrhythmia occurs when the heart beats earlier than it should. This is because of the abnormal electrical activity of the ventricles which causes premature contraction of the lower chambers of heart, the ventricles. The premature contraction is followed by a pause as the heart’s electrical system “resets” itself. The contraction following the pause is usually more forceful than normal. With this type, the QRS complex is misshapen and prolonged representing ventricular contraction without earlier atrial stimulation (see Figure 2.9).
6- Atrial Premature Beat (A): this arrhythmia is associated with early depolarisation of atrium This type can be identified by a premature, small and distorted P-wave (see Figure 2.10).
7- Aberrated Atrial Premature Beat (a): early depolarisation of atria. This manifests itself as an abnormal P-wave (wide prolonged), narrow R-wave, and distorted QRS complex (see Figure 2.11).
8- Nodal (junctional) Escape Beat (j): the cause of this arrhythmia is that the region around the AV node takes over as the focus of the depolarisation; the rhythm is called “nodal” or ‘junctional’ escape. Figure 2.12 shows one beat cycle of this arrhythmia which has no Q- and S-waves. Also, the P-wave has an inverse polarity compared to that of the normal sinus rhythm
potential /mV 0 .7 5 0 .5 0 0 .2 5 0 -0 2 5 -0 .5 0 -0 .7 5 0.6 0.2 0 .4 0 Time/seconds
Figure 2.8: Beat stimulated by an artificial pacemaker ( ‘Pace’) type (MIT-BEH database, record 104)
Electrical potential (mV) 1.25 1.0 0 .7 5 0 .5 0 .2 5 -0 .2 5 -0 .5 0 .4 0.2 0 0.6 Time (seconds)
Figure 2.9: Premature ventricular contraction (V) type
potential (mV) 0 .7 5 0 .5 0 .2 5 -0 .2 5 -0 .5 -0 .7 5 0.8 0.6 0 0.2 0 .4 Time (seconds)
Figure2.10: Atrial premature beat (A) type
Electrical potential (mV) 1.0 0 .7 5 0 .5 0 0 .2 5 -0 .2 5 - 0 .5 0 -0 .7 5 0.2 0.8 0 0 .4 0.6 Time (seconds)
Figure 2.11: Aberrated atrial premature beat (a) type
(MIT-BIH database, Record 105)
potential (mV) 1.5 1.0 0 .5 0 -0 .5 -1.0 -1 .5 0.2 0 .4 0.6 0.8 Time (seconds)
Figure 2.12: Nodal (junctional) escape beat (j) type
(MIT-BIH database, Record 201)
9- Ventricular Escape Beat (E): this most commonly occurs when the ventricle contracts without nodal stimulation. This is classically associated with complete heart blockage. The QRS complexes are wide whereas the P-waves are occasionally absent as demonstrated in Figure 2.13.
10- Fusion of paced and normal beats (f): this type of arrhythmia is a mixture of paced and normal beats. The P-waves have large amplitudes and are wide, and the QRS complexes are distorted, especially in the S-waves portion (see Figure 2.14).
Examples of the above arrhythmias and normal ECGs were extracted from records 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 111, 112, 113, 114, 115, 116, 117, 118, 119, 121, 122, 123, 124, 200, 201, 202, 203, 205, 207, 208, 209, 210, 212, 213, 214, 215, 217, 219, 220, 221, 222, 223, 228, 230, 231, 232, 233, 234.
There are two points to be taken into account concerning the above examples: intra patient and inter-patient variability. Intra-patient variability occurs due to changes in the patient’s emotional and physical states and inter-patient variability is due to different physical conditions between the different patients. As a result of intra- and inter-patient variability, different beat waveforms and different lengths of a beat cycle are observed.
potential (mV) 0 .7 5 0 .5 0 0 .2 5 -0 .2 5 -0 5 0 -0 .7 5 0 0.2 0 .4 0.6 0.8 Time (seconds)
Figure 2.13: Ventricular escape beat (E) type
Electrical potential (mV) 0 .7 5 0 .5 0 0 .2 5 -0 .2 5 -0 .5 0 -0 .7 5 0 0.2 0 .4 0.6 0.8 Time (seconds)
Figure 2.14: Fusion of paced and normal beats (f)
2.4 Data Preparation
The application of a pattern classifier first requires the selection of features that must be tailored separately for each problem domain. Features should contain the information required to distinguish between classes, be insensitive to irrelevant variability in the input, and also limit the amount of training data required. Good classification performance requires selection of effective features and also selection of a classifier that can make good use of those features without demanding large amounts of training data, memory, and computing power [Lippmann, 1989].
Raw ECG data cannot be directly used for classification; it needs to undergo pre processing operations such as filtering, digitisation, feature extraction and normalisation. The MIT-BIH Institute, which produced the ECG database used in this study, had carried out the filtering and digitisation.
The next section explains the preparation of ECG data in terms of feature extraction and normalisation.
2.4.1 Feature Extraction
Feature extraction techniques try to reduce the amount of data to be processed by a pattern classifier by extracting important features that can be used to represent the whole data set. Primary features are directly extracted from a data set. Additional features can be obtained by combining primary features by means of mathematical operations, for example taking the difference or the ratio of two feature values.
The MIT-BIH database consists of records, each pertaining to several ECG beats, with each digitised beat comprising hundreds of points representing one cycle of the ECG. It is very important to reduce these large data files to small files that retain sufficient information to enable differentiating between the different types of arrhythmia. The main problem when classifying an ECG is to determine which portion of the signal to use for diagnosis. Different portions and various numbers of features can be adopted. Many researchers use the QRS-complex as it represents ventricular depolarisation, and because it contains most of the information about the nature of disease [Conde, 1994; Suzuki, 1995; Hosseini and Nazeran, 1999; Biel, 2001]. However, the ST segment should also be taken into account as well as beat rhythm information [Weisner et al, 1982].
It has been noted by Suzuki [Suzuki, 1995] that since the QRS complex is a reference used to detect other waves, the first step in an automatic ECG interpretation system is recognition of this QRS complex. It is also true that the QRS complex contains a significant amount of information about the state of the heart, as it represents ventricular depolarisation [Biel, 2001].
Various sets of features were extracted for this study. In the work reported in chapter 3, two sets of extracted features (one with 18 and the other with 11 features) were used. Another set of 15 features was employed in the experiments reported in chapters 4 and 5. Figure 2.15 shows some of the features extracted from one ECG signal.
RT PR QT QR | RSi QRS ST PQ
Figure 2.15: Some features extracted from an ECG
These features include some of the important information about the ECG signal such as the height and duration of different parts of the waveform Table 2.3 and Table 2.4 respectively, show the 18 and 11 features used in the work described in chapter 3. The second set of features was obtained by removing seven less important features from the first set.
The process of physiological feature extraction starts with the identification of the R- wave, as the QRS complex is the most prominent part of the ECG as mentioned before. The R-wave is normally positive and generally shows a significantly greater peak than the other waves and the R peak can be detected as the highest value in the cycle.
However, the R-wave can sometimes be negative if the electrodes are attached with reversed polarity, and an inverted waveform is seen; in this case the Q- and S-waves are shown as positive.
On some occasions, the T-wave can have a greater magnitude than the R-wave. In this case, distinction is still possible because the R-wave has a pointed peak whereas T- waves are wider and more rounded.
As for the Q-wave and R-wave peaks, they may then be found by searching for a local maximum or minimum In the case of the Q-wave, the search proceeds from the peak of the R-wave towards the left (i.e. the search is made backwards in time). For the S- wave, the search proceeds forward in time from the location of the R-wave peak.
Feature Description
1- P peak height The amplitude of the P-wave.
2- Q peak height The amplitude of the Q-wave.
3- R peak height The amplitude of the R-wave.
4- S peak height The amplitude of the S-wave.
5- T peak height The amplitude of the T-wave.
6- PT wave duration Overall duration of the P & T-wave.
7- PR wave duration Overall duration of the P & R-wave.
8- RT interval Overall duration of the R & T-wave.
9- QRS interval Overall duration of the QRS complex: from the
onset of the Q-wave to the end of the S-wave. (Time taken for complete ventricular pumping action.)
10- QT interval Overall duration of the Q & T-wave.
11- QR interval Overall duration of the Q & R-wave.
12- Minimum The minimum value of the ECG signal
13- Maximum The maximum value of the ECG signal
14- RS interval Overall duration of the R & S-wave.
15- PQ interval Overall duration of the P & Q-wave.
16- ST interval Overall duration of the S & T-wave.
17- Standard deviation Standard deviation of the electrical signal from the Baseline
18- Mean Mean of the electrical signal
Table 2.3: 18 Features of the ECG signal selected for classification
Feature
Description
1- PT wave duration Overall duration of the P & T-wave.
2- PR wave duration Overall duration of the P & R-wave.
3- RT interval Overall duration of the R & T-wave.
4- QRS interval Overall duration of the QRS complex: from the
onset of the Q-wave to the end of the S-wave. (Time taken for complete ventricular pumping action.)
5- QT interval Overall duration of the Q & T-wave.
6- QR interval Overall duration of the Q & R-wave.
7- Minimum The minimum value of the ECG signal
8- Maximum The maximum value of the ECG signal
9- RS interval Overall duration of the R & S-wave.
10- PQ interval Overall duration of the P & Q-wave.
11- ST interval Overall duration of the S & T-wave.
Table
2.4: 11 Features of the ECG signal selected for classificationMIT-BIH database even for the normal type. Some variations in the QRS complex were taken into account during the feature selection process see Figures (2.16(a), (b) and (c)), (2.17(a), (b) and (c)) and (2.18(a), (b), and (c)).
As shown in Table 2.3 and Table 2.4, a number of statistical features was also used.
Another method of reducing the data set used to describe the ECG was to re-sample at a reduced sampling frequency. This method is simpler and faster than the feature extracting methods.
For the experiments reported in chapter 3, two sets of re-sampled data were used. The re-sampling frequencies were 100 Hz and 50 Hz for the two sets respectively, giving 64 points or 33 points for one cycle of the ECG signal.
2.4.2 Normalisation
As mentioned previously, during collection different equipment was employed with different electrical gains for digitisation of the various records. Consequently, the signals required normalisation before use, and for convenience in neural network training, the data was normalised before being presented to the neural network.
Usually, the data would be normalised between 0 and 1 or between -1 and 1. The main advantage of normalisation is to eliminate the effects of different scales and ranges. In this study, the data was normalised in the range [-1, 1]:
s
Qo
R
Figure 2.16 (a): Variations in the R-wave part of QRS complex showing inverted R-wave
Figure
S-wave has greater magnitude than R-wave
R
0
S
Figure 2.16 (c): Variations in the R-wave part of QRS complex showing apparent double R peak
0
Q
s
Figure 2.17 (a): Variations in the Q-wave part of QRS complex showing depressed PR segment
R
0
S
Figure 2.17 (b): Variations in the Q-wave part of QRS complex showing no Q-wave present
0
Q
Figure 2.17 (c): Variations in the Q-wave part of QRS complex showing indeterminate onset to Q-wave
R
0
S
Figure 2.18 (a): Variations in the S-wave part of QRS complex showing indeterminate end to S-wave
0
Q
s
Figure 2.18 (b): Variations in the S-wave part of QRS complex showing depressed ST segment
R
0
Q
Figure 2.18 (c) Variations in the S-wave part of QRS complex showing apparent separation of S-wave from QR part of complex
(F - F )
V ' max min / /min / J
where Fu is the unsealed value, Fmin the minimum value in the data set, and Fmax the maximum value in the data set.
Another effective normalisation method is to calculate the mean and the standard deviation of the attributes first, and then divide the difference between each attribute value and the mean by the standard deviation:
where X is the mean and <j) the standard deviation of the data s e t , Fu is the unsealed value of the feature and Fs is the new scaled value of the feature.
2.5 Previous Work on ECG/Arrhythmia Classification
Several authors have looked at ECG arrhythmia classification using different means such as statistical methods, expert systems, and supervised neural networks.
Automated interpretation of ECGs began more than 35 years ago [Pibperger and Stallman 1962; Pordy et al., 1968]. Since that time there has been continuous development of expert systems for automated interpretation of ECGs. Automated interpretation of ECGs includes three basic approaches. The first is based on decision logic where a rule-based expert system is used to mimic the decision processes of a cardiologist. The second approach utilises multivariate statistical pattern recognition
F. = (2.2)
to solve a pattern recognition problem [Klingeman and Pipberger, 1967]. The third approaches employing neural networks [Rasiah and Attikiouzel, 1994] and machine learning [Oates et al., 1988] have also been developed.
Over the last few years automated interpretation of ECGs has been widely used as decision support for physicians, with the best interpretation programs performing almost as well as humans. Recent papers have shown that neural networks may be used to improve automated ECG interpretation for myocardial infarction.
Neural networks have been utilised with positive results in various medical diagnoses [Gallant, 1988; Frenster, 1990; Peng and Reggia, 1989]. In computerised ECG, the developed applications have concentrated mainly on beat and diagnostic classification [Gallant, 1988; Degani and Bortolan, 1990; Yeap et al, 1990]. According to Lippmann [Lippmann, 1989], recent interest in neural networks is directed towards practical research. This includes areas of study encompassing pattern recognition and artificial intelligence applications where real-time response is required. Both areas are relevant to ECG classification.
Pedrycz et al. [Pedrycz et al., 1991] used a combination of two pattern recognition techniques, cluster analysis and feed-forward back propagation neural networks, for the diagnostic classification of a 12-lead ECG. The principle of cluster analysis based on the Euclidean distance in parameter space was also applied to the original learning set. The classification accuracy results varied between 51.9% and 84.0% for classifying 7 classes of ECG abnormality.
neural network architectures with supervised and unsupervised learning approaches in performing the automatic analysis of the diagnostic ECG, where seven beat types and 39 features were used. The classification results varied between 91.0% and 94.0% correct classification for all seven types, showing that a classifier based on neural networks can produce a performance at least comparable with those of traditional classifiers. As for the neural network architectures trained with unsupervised techniques, they produced a reasonable classification performance. Interestingly, two additional features used were the age and sex of the subjects. This information is not given in the MIT-BIH database.
A neural network based system, the G N et 2000 ambulatory ECG monitor, was developed by Gamlyn et al. [Gamlyn et al., 1999]. This is a portable, battery-powered unit capable of analysing an ECG in real time. A panel of Kohonen networks is embedded in a 32-bit micro-controller. The system is able to detect variations in the heart rate and P-R interval, changes to the ST segment, ‘ectopic’ beats and certain arrhythmias. Features include 24-hour monitoring and printout of detailed reports. The product is now commercially available.
Hu et al. [Hu et al., 1997] used a patient-adaptable approach to classify ECG beats in the MIT-BIH arrhythmia database. They concentrated on four categories of ECG beats, namely, normal, ventricular premature beat, fusion of normal and ventricular beat and unclassifiable beat. They used a mixture of the Self Organising Feature Map (SOFM) and Learning Vector Quantisation (LVQ) algorithms to develop two expert programs, the global expert program capable of classifying ECG beats from the whole
database and the local expert program, which is a patient-specific expert system The classification accuracy varied between 62.2%-95.9% for different records. The main drawback of the method is the need to create a local expert program for each individual patient.
Edenbrandt et al. [Edenbrandt et al., 1992] used single output MLPs to classify seven different classes of ST-T segments found in the ECG. They used the ST slope and the positive and negative amplitudes of the T-wave as inputs to the MLP. They trained and tested ten MLPs with different configurations of hidden layers and neurons in the hidden layers. The average classification accuracy was between 90.0% and 94.4%.
Izeboudjen and Farah [Izeboudjen and Farah, 1998] proposed an arrhythmia classifier using two neural network classifiers based on the MLP model. The morphological classifier groups the P-waves and QRS complexes into normal or abnormal beats. The timing classifier takes as the input the output of the morphological classifier and the duration of the PP, PR and RR intervals (see Figure 2.19). An accuracy of 93.0% was reported in classifying 13 arrhythmia classes from 48 examples scanned from different ECG signals using a PC.
Dorffher et al. [Dorffner et al., 1994] compared the performance of neural networks with the performance of skilled cardiologists in classifying coronary artery disease during stress and exercise testing. He performed three experiments, two of which used recurrent networks, while the third one employed an MLP. This neural network approach produced results comparable to the diagnosis of experts. Only in some cases did the neural networks outperform the experts.
RR interval Voltage, mV R ;! f\
^ 1
A
I ’ "*£*-R R < i / y \ ! w y I K-iry
Time, SecondsNugent et al. [Nugent at al., 1998] used single-output bi-group MLPs to detect the presence or absence of a specific ECG class. Three different feature selection techniques were adopted, namely, rule based, manual and statistical. The results of the bi-group neural networks were combined using orthogonal summation. The methodology was applied to recognise three classes, namely, normal, left ventricular hypertrophy and inferior myocardial infarction. On average, the classification accuracy was only 78.0%.
Biel et al. [Biel et al., 2001] suggested that the distinction between ECG signals of different people is sufficiently great to identify individuals using just one lead of an ECG.
Bortolan et al. [Bortolan et al., 1991] used a feed forward network with backpropagation to classify seven beat types using 39 features. Results of over 90.0% correct classification for all seven types were achieved. Interestingly, two features used were the age and sex of the subjects. Such information is not given in the MIT- BIH database. The same seven beat classes were investigated by Silipo et al. [Silipo et al., 1999] using a neural classifier with Radial Basis Function (RBF) pre-processing. Here again, correct beat type designation was consistently made for over 90.0 % for all classes.
The influence of various network parameters on multilayer neural network performance were researched by Edenbrandt et al. [Edenbrandt et al., 1992]. ECG ST- T segments were the basis of the study which found that increasing the number of input features did not necessarily improve classification. Similarly, increasing the
reported that networks with two hidden layers showed only a very slight improvement over those with one hidden layer. Problems were encountered with training networks to recognise uncommon patterns, the best results being obtained, as expected, for those beats with the most examples in the training set.
Magleveras et al. [Magleveras et al., 1998] advised against using digital filtering of signals at the pre-processing stage to avoid corrupting the components of the ECG. However, others have done so in their work [Hamilton et al., 1986; Suzuki, 1995; Dokur et al., 1997].
Modular neural networks were applied to ECG classification [Kidwai, 2001]. These employed a more logical step-by-step approach by breaking the problem of classification down into stages rather than using a one-hit approach.
Suzuki [Suzuki, 1995] and Hamilton and Tompkins [Hamilton and Tompkins, 1986] researched methods of QRS complex detection. Their aim was reliably to break down a continuous ECG signal into individual beats. This is in contrast to supplying information from a database where signals have already been pre-divided into beats, such as the MIT-BIH database. Recognition of the QRS complex was proposed by Suzuki as the first step in the development of a real-time ECG analysis system His self-organising neural network was capable of detecting R-waves in real time, in order to divide the ECG into cardiac cycles. An Adaptive Resonance Theory (ART) network then performed classification according to QRS complex features. Hamilton and Tompkins [Hamilton and Tompkins, 1986] claimed that their system carried out
